The vertical deflection of the main girder on a cable-stayed bridge is a direct reflection of the vertical stiffness of bridge structure, which represents the comprehensive mechanical performance of cable-stayed bridge. Compared with the deflection caused by vehicles, the deflection caused by temperature is often more significant and the change frequency is very low, which is easy to extract from raw data, and can be used as an index to evaluate the state of cable-stayed bridge. To obtain the control value for recognizing the abnormal deflection, it is necessary to establish an accurate input-output relationship between temperature and temperature- induced deflection. However, because of the high-order nonlinear relationship between the temperature and the temperature-induced deflection, the traditional linear regression is not accurate enough in modeling this relationship. To establish a high-precision model for the deflection, this paper uses the machine learning tools with the highly nonlinear fitting performance to further model the project. Considering both the precision and modeling efficiency, the Long-Short Term Memory (LSTM) network can build the optimal model between temperature and temperature-induced deflection. Use the regression value output by LSTM as the control value combining with the statistical pattern of t-test, the 6% abnormal deflection can be recognized. The 6% sensitivity can help to recognize bridge abnormalities earlier.